Concept drift
The relationship between inputs and the right answer changing, so a model can be wrong on data that looks entirely familiar. Fraud patterns adapting to a detector is the classic case.
What people get wrong
It cannot be detected by watching the inputs, which is what distinguishes it from data drift. It only shows up once outcomes arrive, so monitoring has to include a way of collecting them.
A definition is not an understanding. This is taught properly in Keeping a model working after launch, part of MLOps & model lifecycle.
Learn this properlyTerms from the same topic